README.md

---
license: mit
task_categories:
  - visual-question-answering
  - question-answering
language:
  - en
tags:
  - robotics
  - 6dof-pose
  - grasping
  - spatial-reasoning
  - trajectory
  - depth-estimation
  - benchmark
  - evaluation
size_categories:
  - n<1K
pretty_name: BOPASK-Test
configs:
  - config_name: core-handal
    data_files:
      - split: test
        path: core/bopask-test-handal.json
  - config_name: core-hope
    data_files:
      - split: test
        path: core/bopask-test-hope.json
  - config_name: core-ycbv
    data_files:
      - split: test
        path: core/bopask-test-ycbv.json
  - config_name: lab-home
    data_files:
      - split: test
        path: lab/bopask-test-home.json
---

# BOPASK-Test

Human-verified evaluation benchmark for the **BOPASK** spatial-reasoning VQA dataset.

Contains **934 question-answer pairs** across **two testsets**:

- **`core`** — BOPASK-Core: three BOP-Challenge families (HANDAL, HOPE, YCB-V).
- **`lab`**  — BOPASK-Lab : an in-the-wild set of "home / lab" scenes.

## Contents at a glance

| Split | Family  | Records | RGB images | Depth maps | Masks |
|-------|---------|--------:|-----------:|-----------:|------:|
| core  | handal  |     251 |         43 |         41 |   138 |
| core  | hope    |     189 |         50 |         29 |   231 |
| core  | ycbv    |     248 |         48 |         48 |   153 |
| lab   | home    |     246 |         21 |         12 (⚠) |    52 |
| **Total** |     |  **934** |    **162** |    **130** | **574** |


## Question-type distribution

| question_type / subtype | handal | hope | ycbv | home | **Total** |
|---|---:|---:|---:|---:|---:|
| pose / 2dbbox | 39 | 39 | 38 | 39 | **155** |
| grasp / 2dplane | 40 | 40 | 40 | 38 | **158** |
| spatial_reasoning / relative_position | 40 | 40 | 40 | 71 | **191** |
| trajectory / 2d | 40 | 40 | 40 | 48 | **168** |
| depth_relative / closer | 40 | — | 40 | 16 | **96** |
| depth_relative / farther | 40 | — | 40 | 24 | **104** |
| object_rearrangement / point_wise | 12 | 30 | 10 | 10 | **62** |
| **family total** | **251** | **189** | **248** | **246** | **934** |

## Layout

```
bopask-test/
├── README.md
├── core/                           (BOPASK-Core testset)
│   ├── bopask-test-handal.json
│   ├── bopask-test-hope.json
│   ├── bopask-test-ycbv.json
│   ├── handal/
│   │   ├── images/                 (43 *.png)
│   │   ├── depth_maps/             (41 *_depth.png)
│   │   └── masks/                  (138 *_mask.png)
│   ├── hope/
│   │   └── images/  depth_maps/  masks/
│   └── ycbv/
│       └── images/  depth_maps/  masks/
└── lab/                            (BOPASK-Lab testset)
    ├── bopask-test-home.json
    └── home/
        ├── images/                 (21 *.png)
        ├── depth_maps/             (empty — see caveat above)
        └── masks/                  (52 masks_<scene>_<object>.png)
```

All paths inside each JSON are **relative to this dataset root**, e.g.
`core/handal/images/scene_000008_frame_000980.png`.

## Quick start

```python
import json
from datasets import load_dataset

# Load one of the configs:
ds = load_dataset("bhatvineet/bopask-test", "core-handal", split="test")
print(ds[0])

# Or load all four families manually:
configs = ["core-handal", "core-hope", "core-ycbv", "lab-home"]
for cfg in configs:
    d = load_dataset("bhatvineet/bopask-test", cfg, split="test")
    print(cfg, len(d))
```

Loading directly without `datasets`:

```python
import json
with open("core/bopask-test-handal.json") as f:
    records = json.load(f)

for r in records:
    img_path  = r["images"][0]          # e.g. "core/handal/images/..."
    user_q    = r["messages"][0]["content"]
    gt_answer = r["messages"][1]["content"]
```

## Evaluation protocols

Each record is a single-turn VQA pair with one ground-truth response in
`messages[1].content`. Answer formats are self-describing — the user prompt
tells the model the expected output format (e.g. "respond as a list of 2D
points…"). Common metrics by type:

| question_type | typical metric |
|---|---|
| pose / 2dbbox | 2D IoU |
| grasp / 2dplane | endpoint L2 / success@τ |
| trajectory / 2d | trajectory-wise DTW, endpoint error |
| spatial_reasoning / relative_position | exact match (yes/no) |
| depth_relative | exact match (closer/farther) |
| object_rearrangement / point_wise | point-in-mask accuracy |

## Relationship to the training set

This benchmark was curated and human-verified to be disjoint from the
[`bhatvineet/bopask-train`](https://huggingface.co/datasets/bhatvineet/bopask-train)
training split. Use this for evaluation only.

## Citation

If you use this dataset, please cite the [BOPASK](https://arxiv.org/abs/2511.16857) paper
and the underlying BOP-Challenge object-pose datasets
(HANDAL, HOPE, LineMOD, YCB-V).

## License

MIT for the QA annotations. The underlying RGB / depth / mask assets inherit
the licenses of their source BOP-Challenge datasets (HANDAL, HOPE, YCB-V) and
the bopask-home captures.